Journal of Health Economics 2012 Cawley

Contents

lists

available

at

SciVerse

ScienceDirect

Journal

of

Health

Economics

j

o u r n a

l

h o

m e

p a g e :

w w w . e l s e v i e r . c o m / l o c a t e / e c o n b a s e

The

medical

care

costs

of

obesity:

An

instrumental

variables

approach

!

John

Cawley

a

,

b

,

c

,

∗

,

Chad

Meyerhoefer

d

a

_Department

_of

_Policy

_Analysis

_and

_Management,

_Cornell

_University,

_United

_States

b

_Department

_of

_Economics,

_Cornell

_University,

_United

_States

c

_NBER,

_United

_States

d

_Department

_of

_Economics,

_Lehigh

_University,

_United

_States

a

r

t

i

c

l

e

i

n

f

o

Article

history:

Received

7

October

2010

Received

in

revised

form

29

September

2011

Accepted

11

October

2011

Available online 20 October 2011

JEL

classification:

I1

I14

I18

D6

Keywords:

Obesity

Externalities

Medical

care

costs

Insurance

a

b

s

t

r

a

c

t

This

paper

is

the

first

to

use

the

method

of

instrumental

variables

(IV)

to

estimate

the

impact

of

obesity

on

medical

costs

in

order

to

address

the

endogeneity

of

weight

and

to

reduce

the

bias

from

reporting

error

in

weight.

Models

are

estimated

using

restricted-use

data

from

the

Medical

Expenditure

Panel

Sur-vey

for

2000–2005.

The

IV

model,

which

exploits

genetic

variation

in

weight

as

a

natural

experiment,

yields

estimates

of

the

impact

of

obesity

on

medical

costs

that

are

considerably

higher

than

the

estimates

reported

in

the

previous

literature.

For

example,

obesity

is

associated

with

$656

higher

annual

medical

care

costs,

but

the

IV

results

indicate

that

obesity

raises

annual

medical

costs

by

$2741

(in

2005

dollars).

These

results

imply

that

the

previous

literature

has

underestimated

the

medical

costs

of

obesity,

result-ing

in

underestimates

of

the

economic

rationale

for

government

intervention

to

reduce

obesity-related

externalities.

© 2011 Elsevier B.V. All rights reserved.

1.

Introduction

In

the

United

States,

the

prevalence

of

obesity,

defined

as

a

body

mass

index

1

_or

_BMI

_>

_30,

_has

_been

_rising

_for

_at

_least

_five

_decades

(e.g.

Burkhauser

et

al.,

2009;

Komlos

and

Brabec,

2010

)

and

has

more

than

doubled

in

the

past

thirty

years

(

Flegal

et

al.,

1998

).

In

2007–2008,

33.8%

of

American

adults

were

clinically

obese

(

Flegal

et

al.,

2010

).

This

is

troubling

because

obesity

is

associated

with

an

increased

risk

of

myocardial

infarction,

stroke,

type

2

diabetes,

can-cer,

hypertension,

osteoarthritis,

asthma,

and

depression,

among

other

conditions

(

Dixon,

2010;

Hu,

2008

).

!

_We

_thank

_Tania

_Andreyeva,

_Virginia

_Chang,

_Eric

_Finkelstein,

_Alan

_Monheit,

_Leo

Trasande,

Jessica

P.

Vistnes,

and

two

anonymous

referees

their

helpful

comments.

We

are

particularly

indebted

to

Joe

Newhouse,

who

provided

many

thoughtful

and

detailed

suggestions.

∗

Corresponding

author

at:

Department

of

Policy

Analysis

and

Management,

Cor-nell

University,

Ithaca

NY

14853,

United

States.

Tel.:

+1

607

255

0952;

fax:

+1

607

255

4071.

E-mail

addresses:

JHC38@cornell.edu

(J.

Cawley),

CHM308@lehigh.edu

(C.

Meyerhoefer).

1

_Body

_mass

_index

_is

_defined

_as

_weight

_in

_kilograms

_divided

_by

_height

_in

_meters

squared.

Many

previous

papers

have

estimated

the

association

of

obesity

with

medical

care

costs

(e.g.

Finkelstein

et

al.,

2009;

Trasande

et

al.,

2009;

Thorpe

et

al.,

2004;

Finkelstein

et

al.,

2003;

Kortt

et

al.,

1998

).

Typically,

this

involves

estimating

cross-sectional

models

using

large

secondary

datasets

such

as

the

National

Medical

Expenditure

Survey

of

1987

(NMES)

and

the

more

recent

Medical

Expendi-ture

Panel

Survey

(MEPS).

These

studies

have

made

an

important

contribution

to

the

literature

by

demonstrating

the

significance

of

medical

costs

associated

with

obesity

and

the

diseases

linked

to

obesity.

As

a

result,

these

papers

have

been

heavily

cited

and

widely

influential.

2

_For

_example,

_these

_estimates

_have

_been

_used

_to

jus-tify

government

programs

to

prevent

obesity

on

the

grounds

of

external

costs

(e.g.

U.S.

D.H.H.S.,

2010

).

However,

the

previous

estimates

have

important

limitations.

The

most

significant

is

that

they

measure

the

correlation

of

obe-sity

with,

not

the

causal

effect

of

obesity

on,

medical

care

costs.

The

correlation

is

an

overestimate

of

the

causal

effect

if,

for

exam-ple,

some

people

became

obese

after

suffering

an

injury

or

chronic

depression,

and

have

higher

medical

costs

because

of

the

injury

or

2

_For

_example,

_Finkelstein

_et

_al.

₍₂₀₀₃₎

_has

_been

_cited

₂₃₅

_times,

_as

_of

_September

9,

2011,

according

to

the

ISI

Web

of

Knowledge.

0167-6296/$

–

see

front

matter

©

2011 Elsevier B.V. All rights reserved.

depression

(which

is

likely

to

be

unobserved

by

the

econometri-cian).

Conversely,

the

correlation

is

an

underestimate

of

the

causal

effect

if,

for

example,

those

with

less

access

to

care,

such

as

dis-advantaged

minorities

and

the

poor,

are

more

likely

to

be

obese

(

Fontaine

and

Bartlett,

2000

).

Another

limitation

is

that

these

stud-ies

are

usually

based

on

self-reported,

rather

than

measured,

height

and

weight,

and

this

reporting

error

biases

the

coefficient

estimates

(

Bound

et

al.,

2002

).

This

paper

builds

on

the

previous

research

by

addressing

both

of

these

problems

–

endogeneity

of

weight

and

reporting

error

in

weight

–

by

estimating

models

of

instrumental

variables.

Our

instrument

for

the

respondent’s

weight

is

the

weight

of

a

biological

relative,

an

instrument

used

in

the

previous

literature

to

estimate

the

impact

of

weight

on

other

outcomes

such

as

wages

(e.g.

Cawley,

2004;

Kline

and

Tobias,

2008

)

and

mortality

(

Smith

et

al.,

2009

).

We

estimate

the

IV

model

using

the

2000–2005

MEPS,

the

lead-ing

source

of

data

on

medical

care

costs

and

utilization

for

the

U.S.

non-institutionalized

population.

Our

results

indicate

that

the

effect

of

obesity

on

medical

care

costs

is

much

greater

than

previ-ously

appreciated.

The

model

also

passes

several

falsification

tests:

it

finds

a

stronger

impact

of

obesity

on

medical

expenditures

for

diabetes

(clearly

linked

to

obesity)

than

on

medical

expenditures

for

other

conditions,

does

not

find

an

impact

of

obesity

on

medical

care

costs

for

conditions

that

are

unrelated

to

obesity,

and

bio-logically

unrelated

children

(e.g.

stepchildren)

are

not

significant

predictors

of

respondent

weight.

The

limitations

of

cost

of

illness

studies

are

widely

recognized

(

Shiell

et

al.,

1987;

Roux

and

Donaldson,

2004

).

For

example,

they

are

not

useful

for

prioritizing

the

allocation

of

medical

resources

because

that

would

amount

to

a

circular

argument:

some

condi-tions

have

a

large

amount

of

resources

devoted

to

them

and

thus

have

a

high

cost

of

illness,

but

that

does

not

imply

that

even

more

funding

is

needed

(see,

e.g.,

Shiell

et

al.,

1987

).

This

paper

does

not

estimate

the

medical

care

costs

of

obesity

in

order

to

argue

that

treatment

of

obesity

should

be

prioritized

above

treatment

of

other

conditions,

but

to

more

accurately

measure

the

marginal

effect

of

obesity

on

medical

care

costs.

2.

Empirical

model

2.1.

Identification:

method

of

instrumental

variables

Ideally,

to

measure

the

effect

of

obesity

on

medical

care

costs

one

would

conduct

a

randomized

controlled

trial

in

which

obesity

was

assigned

by

the

investigator.

Such

an

experiment

would,

of

course,

be

unethical,

so

one

must

rely

on

natural

experiments.

We

follow

the

previous

literature

(e.g.

Cawley,

2004;

Kline

and

Tobias,

2008;

Smith

et

al.,

2009

)

and

use

the

weight

of

a

biological

relative

as

an

instrument

for

the

weight

of

the

respondent.

There

are

two

requirements

for

an

instrument.

First,

it

must

be

powerful.

The

weight

of

a

biological

relative

is

a

powerful

pre-dictor

of

the

weight

of

a

respondent

because

roughly

half

the

variation

in

weight

across

people

is

genetic

in

origin

(

Comuzzie

and

Allison,

1998

).

As

we

describe

in

Section

4

,

our

instrument

set

easily

exceeds

the

conventional

benchmark

for

power

of

F

=

10

in

the

first

stage

(

Stock

et

al.,

2002

).

The

second

requirement

is

valid-ity

–

the

instrument

must

be

uncorrelated

with

the

error

term

in

the

second

stage.

In

the

present

context,

this

means

that

the

weight

of

a

biological

relative

must

be

uncorrelated

with

the

respondent’s

residual

medical

care

costs

after

controlling

for

predicted

respon-dent

weight

and

other

observed

characteristics.

Validity

would

be

threatened

if

both

the

respondent

and

the

biological

relative

are

affected

by

a

common

household

environ-ment

that

is

also

directly

correlated

with

the

respondent’s

medical

expenditures.

Although

it

is

impossible

to

prove

the

null

hypothe-sis

of

no

effect,

and

therefore

some

doubt

will

always

remain,

much

research

in

behavioral

genetics

finds

no

detectable

effect

of

a

shared

household

environment

effect

on

weight.

Adoption

studies

have

consistently

found

that

the

correlation

in

weight

between

a

child

and

its

biological

parents

is

the

same

for

children

raised

by

their

biological

parents

and

children

raised

by

adoptive

parents

(

Vogler

et

al.,

1995;

Stunkard

et

al.,

1986;

Sorensen

and

Stunkard,

1993

).

Other

studies

have

found

that

the

weights

of

unrelated

adopted

sib-lings

are

uncorrelated

(

Grilo

and

Pogue-Geile,

1991

).

Twin

studies

(which

by

necessity

are

based

on

small

samples)

find

no

signif-icant

difference

between

the

correlation

in

the

weight

of

twins

reared

together

and

twins

reared

apart

(

Price

and

Gottesman,

1991;

Maes

et

al.,

1997

),

which

is

consistent

with

a

negligible

common

household

environment

effect

on

weight.

With

hundreds

of

behavioral

genetics

studies

on

the

subject,

there

are

of

course

some

studies

that

detect

a

shared

family

envi-ronment

on

BMI

(e.g.

Nelson

et

al.,

2006

),

but

the

preponderance

of

evidence

is

that

any

such

effects

are

so

small

as

to

be

undetectable

and

ignorable

(

Hewitt,

1997;

Grilo

and

Pogue-Geile,

1991;

Maes

et

al.,

1997

).

For

example,

a

recent

study

using

the

same

data

as

Nelson

et

al.

(2006)

concluded:

“We

also

did

not

find

any

support

for

shared

environmental

effects

on

BMI

at

any

age.”

(

Haberstick

et

al.,

2010

,

p.

501).

This

may

be

contrary

to

conventional

wisdom

but

it

is

a

robust

finding;

a

comprehensive

review

concluded

that

“[E]xperiences

that

are

shared

among

family

members

appear

largely

irrelevant

in

determining

individual

differences

in

weight

and

obesity”

(

Grilo

and

Pogue-Geile,

1991

),

and

more

recently

Wardle

et

al.

(2008)

con-clude:

“Contrary

to

widespread

assumptions

about

the

influence

of

the

family

environment,

living

in

the

same

home

in

childhood

appears

to

confer

little

similarity

in

adult

BMI

beyond

that

expected

from

the

degree

of

genetic

resemblance.”

(

Wardle

et

al.,

2008

,

p.

398)

One

must

always

be

cautious

with

regard

to

the

validity

of

instruments,

but

given

the

consistent

finding

that

similarity

in

weight

between

biological

relatives

can

be

attributed

to

genet-ics,

we

believe

that

there

is

enough

suggestive

evidence

regarding

power

and

validity

to

proceed

with

the

use

of

weight

of

a

biologi-cal

relative

as

an

instrument

for

respondent

weight.

As

a

check

of

validity,

we

later

conduct

a

falsification

test

that

uses

the

weight

of

a

stepchild

(when

available)

instead

of

a

biological

child

and

find

that

the

weight

of

a

stepchild

is

not

a

significant

predictor

of

respondent

weight,

which

is

consistent

with

our

identifying

assumption.

In

the

previous

literature

on

the

medical

care

costs

of

obe-sity,

coefficients

are

likely

biased

because

of

measurement

error

in

BMI

that

is

due

to

using

self-reported,

rather

than

measured,

weight

and

height.

3

_(Only

_self-reports

_or

_{proxy-reports}

_of

_weight

and

height

are

available

in

the

MEPS.)

Numerous

studies

have

documented

systematic

misreporting

of

height

and

weight

(e.g.

Plankey

et

al.,

1997;

Villanueva,

2001

).

For

example,

Cawley

and

Burkhauser

(2006)

examine

data

from

the

National

Health

and

Nutrition

Examination

Survey

III,

which

contains

data

on

both

self-reported

and

measured

weight

and

height.

Using

self-reported,

rather

than

measured,

data

to

calculate

BMI

results

in

consider-able

underestimation

of

the

prevalence

of

obesity;

e.g.

among

white

females,

the

prevalence

of

obesity

is

21.6%

based

on

measurements

221

but

17.4%

based

on

self-reports

(

Cawley

and

Burkhauser,

2006

).

Another

source

of

reporting

error

is

that

much

of

the

information

contained

in

the

MEPS

is

reported

by

a

single

household

member,

or

proxy,

and

it

is

possible

that

proxies

may

not

provide

accurate

height

and

weight

information

for

others

in

the

household.

(On

the

other

hand,

it

is

also

possible

that

proxies

report

other

peo-ple’s

weight

more

accurately

than

respondents

report

their

own

weight.)

Both

of

these

sources

of

reporting

error

are

expected

to

bias

the

coefficient

estimates

in

the

previous

literature.

A

benefit

of

the

IV

method

is

that

it

corrects

for

these

multiple

sources

of

measurement

error

(see,

e.g.

Bound

et

al.,

2002

).

2.2.

Two-part

model

of

medical

expenditures

To

estimate

the

impact

of

BMI

and

obesity

on

medical

spending

we

use

a

two-part

model

(2PM)

of

medical

expenditures

(

Jones,

2000

).

The

first

part

of

the

2PM

estimates

the

probability

of

positive

medical

expenditures,

while

the

second

part

estimates

the

amount

of

medical

expenditures

conditional

on

having

any.

We

specify

the

first

part

as

a

Logit

model

and

the

second

part

as

a

Gamma

GLM

with

log

link.

4

_Following

_the

_suggestion

_of

_Manning

_and

_Mullahy

₍₂₀₀₁₎

_,

we

used

modified

Park

tests

to

determine

the

proper

choice

of

the

conditional

variance

function

for

the

GLM,

and

Hosmer–Lemeshow

tests

to

confirm

that

our

choice

of

link

function

is

consistent

with

the

data

generating

process.

5

Given

our

specification

of

the

2PM,

both

parts

of

the

model

require

the

use

of

nonlinear

instrumental

variables

techniques.

Because

both

the

Logit

model

and

the

Gamma

model

are

among

the

class

of

GLMs,

one

can

use

the

instrumental

variable

estimator

of

Carroll

et

al.

(1995)

and

Hardin

et

al.

(2003)

to

determine

the

effect

of

weight

on

medical

expenditures

when

the

endogenous

and

mismeasured

regressor

is

either

BMI

or

a

discrete

indicator

for

obesity.

6

_In

_both

_cases,

_our

_primary

_set

_of

_instruments

_for

_the

4

_Identifying

_the

_appropriate

_functional

_form

_for

_the

_second

_part

_of

_2PM

_requires

analysis

of

various

characteristics

of

the

expenditure

distribution.

An

additional

consideration

in

this

case

is

that

we

seek

to

provide

estimates

for

the

overall

popu-lation

of

non-elderly

adults

as

well

as

seven

sub-populations

(men,

women,

white,

non-white,

private

insurance,

Medicaid,

uninsured),

so

our

estimator

must

perform

well

across

sixteen

different

combinations

of

sample

and

empirical

specification.

The

two

most

widely

used

estimators

for

the

second

part

of

the

2PM

are:

(1)

OLS

of

the

log

of

the

dependent

variable;

and

(2)

the

GLM

estimator.

A

significant

draw-back

of

the

log

OLS

approach

is

that

re-transformation

of

the

estimates

back

to

the

raw

scale

requires

knowledge

of

the

degree

and

form

of

heteroscedasticity.

In

our

application

this

would

entail

the

difficult

task

of

accurately

diagnosing

and

correct-ing

for

heteroscedasticity

in

each

sub-sample,

making

the

GLM

approach

attractive

in

comparison.

However,

GLMs

can

be

inefficient

if

the

log-scale

disturbances

are

heavy-tailed

(Manning

and

Mullahy,

2001),

so

we

examined

the

kurtosis

of

the

log-scale

residuals

from

an

OLS

model

of

medical

expenditures

and

found

it

has

an

average

value

of

3.2

in

our

data.

While

this

is

slightly

larger

than

the

normal

dis-tribution,

a

properly

specified

GLM

model

should

be

reasonably

efficient

under

this

degree

of

skewness.

5

_The

_Park

_tests

_indicated

_that

_the

_conditional

_variance

_is

_proportional

_to

_the

square

of

the

conditional

mean

(

!

ranges

from

1.91–2.06

and

is

precisely

estimated),

which

is

consistent

with

a

gamma-class

model.

To

perform

the

Hosmer–Lemeshow

tests

we

regressed

the

prediction

errors

from

each

model

on

deciles

of

the

distribu-tion

of

predicted

expenditures.

If

the

F

-test

of

coefficients

on

the

decile

indicators

is

jointly

significant

it

indicates

that

the

model

does

not

fit

the

data

well

over

the

dis-tribution

of

predicted

expenditures.

We

rejected

the

null

hypothesis

that

the

decile

coefficients

are

jointly

equal

to

zero

for

only

three

out

of

sixteen

models,

which

suggests

that

the

gamma

model

with

log

link

is

broadly

appropriate.

In

addition,

Hill

and

Miller

(2009)

found

that

this

specification

performed

relatively

well

on

the

1996–2003

sample

of

non-elderly

MEPS

respondents

with

private

insurance.

6

_This

_approach

_incorporates

_a

_linear

_first

_stage,

_which

_is

_most

_appropriate

_when

the

endogenous

and

mismeasured

regressor

is

continuous.

While

it

is

not

uncom-mon

to

estimate

IV

models

with

a

linear

first

stage

when

the

endogenous

regressor

is

discrete,

the

resulting

coefficient

estimate

may

be

biased.

If

the

regressor

suf-fers

only

from

nonclassical

measurement

error

then

the

true

effect

will

generally

lie

between

the

OLS

and

IV

estimates

in

the

case

of

a

simple

univariate

regression

(Black

et

al.,

2000).

When

the

regressor

of

interest

is

both

endogenous

and

mismeasured,

two-part

IV

models

is

the

BMI,

BMI

squared,

and

BMI

cubed

of

the

respondent’s

oldest

biological

child.

7

Prior

research

suggests

that

the

relationship

between

body

weight

and

health

status

is

nonlinear.

In

particular,

mortality

risk

is

somewhat

U-shaped

over

BMI,

with

the

underweight

(BMI

<

18.5)

and

obese

(BMI

≥

30)

facing

higher

mortality

risk

than

the

healthy

weight

(18.5

≤

BMI

<

25)

and

overweight

(25

≤

BMI

<

30)

(

Flegal

et

al.,

2005;

Seidell

et

al.,

1996

).

To

accommodate

nonlinearities

in

the

relationship

between

medical

expenditures

and

weight

sta-tus

we

estimate

a

second

set

of

two-part

IV

models

in

which

the

endogenous

regressors

are

the

respondent’s

BMI

and

BMI

squared.

We

also

estimated

exactly

identified

IV

models

that

include

as

endogenous

regressors

the

respondent’s

BMI,

BMI

squared,

and

BMI

cubed.

Both

specifications

yield

a

similar

pattern

of

expendi-tures

over

BMI,

but

we

report

results

from

the

model

that

includes

BMI

and

BMI

squared

but

not

BMI

cubed

because

the

confidence

intervals

around

predicted

medical

expenditures

are

much

nar-rower.

The

impact

of

obesity

on

medical

expenditures

may

vary

across

the

distribution

of

medical

spending.

To

explore

this

possibility

we

estimate

the

conditional

quantile

treatment

effect

(QTE)

of

obesity

at

different

points

in

the

medical

expenditure

distribution

using

Frolich

and

Melly’s

(2008)

implementation

of

the

IV

estimator

of

Abadie

et

al.

(2002)

.

In

this

case

the

instrument

must

be

discrete,

so

we

use

the

obesity

status

of

the

oldest

biological

child.

8

All

of

our

IV

models

control

for

the

following

regressors:

gen-der,

race/ethnicity

(white,

black,

Hispanic,

other

race),

respondent

age

(indicator

variables

for

whether

age

in

years

is

20–34,

35–44,

45–54,

or

55–64),

education

level

(no

high

school

diploma,

high

school

graduate,

some

college,

bachelor’s

degree

or

higher),

census

region

(northeast,

midwest,

south,

or

west),

whether

the

respon-dent

lives

in

an

MSA,

household

composition

(number

of

household

members

age

0–5

years,

6–17,

18–64,

and

65

or

older),

whether

the

survey

information

was

self-reported

as

opposed

to

proxy-reported,

whether

the

individual

was

employed,

fixed

effects

for

year,

the

gender

of

the

oldest

child,

and

the

age

of

the

oldest

child

Frazis

and

Loewenstein

(2003)

demonstrate

that

the

true

effect

lies

within

bounds

applied

to

the

IV

estimate.

An

alternative

approach

is

to

specify

the

exact

distri-bution

of

both

the

binary

endogenous

regressor

and

the

outcome

variable.

Even

if

the

distributional

assumptions

are

not

correct

and

the

treatment

effect

estimate

is

biased,

this

estimator

may

still

be

preferred

from

a

mean

square

error

standpoint.

Deb’s

(2007)

treatment

effects

gamma

model

provides

an

estimator

of

this

type

that

is

appropriate

for

modeling

skewed

outcomes,

such

as

medical

expenditures.

The

estimation

approach

makes

use

of

simulated

maximum

likelihood

techniques

to

predict

the

impact

of

the

treatment

variable

(obesity),

which

is

assumed

to

follow

a

normal

distribution,

on

an

outcome

variable

(medical

expenditures)

generated

from

a

gamma

distribution

(Deb

and

Trivedi,

2006a,b).

To

test

the

sensitivity

of

our

results

to

an

alternative

estimator

that

explicitly

accounts

for

the

discrete

nature

of

the

endogenous

and

mismeasured

regressor,

we

re-estimated

all

of

our

models

using

the

treatment

effects

gamma

model.

While

the

marginal

effects

of

obesity

we

derived

using

this

approach

were

very

similar

to

those

derived

from

the

approach

of

Carroll

et

al.

(1995)

on

all

samples

except

the

uninsured,

we

prefer

the

method

of

Carroll

et

al.

(1995)

because

it

produced

more

consistent

medical

expenditure

predictions

across

the

full

range

of

the

BMI

distribution.

7

_We

_obtained

_similar

_results

_using

_three

_other

_instrument

_sets:

₍₁₎

_BMI,

_BMI

squared,

and

BMI

cubed

of

the

youngest

child;

(2)

BMI

and

BMI

squared

of

the

youngest

child

and

the

BMI

of

the

second

youngest

child;

(3)

BMI

and

BMI

squared

of

the

oldest

child

and

the

BMI

of

the

second

oldest

child.

All

instrument

sets

have

the

same

theoretical

justification,

but

we

prefer

the

BMI,

BMI

squared,

and

BMI

cubed

of

the

oldest

child

because

there

is

a

higher

response

rate

to

the

height

and

weight

questions

for

older

children.

222

in

months.

9

_For

_subgroup

_analyses

_the

_set

_of

_regressors

_is

_modified

to

drop

irrelevant

control

variables.

3.

Data:

medical

expenditure

panel

survey

(MEPS)

The

medical

expenditure

panel

survey

(MEPS)

is

a

com-prehensive,

nationally

representative

survey

of

the

U.S.

civilian

non-institutionalized

population

that

has

been

conducted

annu-ally

since

1996

and

uses

an

overlapping

panel

design.

Respondents

are

surveyed

about

their

medical

care

use

and

expenditures

over

the

course

of

two

years

through

five

interview

rounds.

In

addition,

information

from

the

household

is

supplemented

by

expenditure

data

collected

directly

from

participants’

medical

service

providers

and

pharmacies

through

a

Medical

Provider

Component.

We

use

data

from

the

2000–2005

waves

of

the

MEPS,

and

convert

medical

expenditures

in

each

year

to

2005

dollars.

We

limit

the

sample

to

adults

between

the

ages

of

20

and

64

with

biological

children

between

the

ages

of

11

years

(132

months)

and

20

years

(240

months),

and

exclude

pregnant

women.

We

use

the

restricted-use

MEPS

data,

which

include

relationship

mappings

for

sample

members

in

the

same

household,

to

identify

biologi-cal

children,

stepchildren,

and

foster

children

and

thus

ensure

that

only

biological

children

are

used

as

instruments.

We

do

not

use

information

on

children

younger

than

11

years

because

rates

of

non-response

for

their

height

and

weight

begin

to

exceed

14%

and

worsen

for

younger

children.

The

weight

and

height

of

each

individ-ual

in

the

household

are

typically

reported

by

a

single

respondent,

most

often

the

wife/mother.

10

_We

_excluded

_eighteen

individu-als

with

implausibly

high

BMIs

(greater

than

80),

as

well

as

two

individuals

with

extremely

high

reported

medical

expenditures

in

excess

of

$292,000,

bringing

our

final

estimation

samples

to

9852

men

and

13,837

women.

We

use

various

measures

of

medical

spending

in

our

empiri-cal

models:

total

medical

expenditures,

expenditures

by

all

third

party

payers

(typically,

public

and

private

insurers),

and

also

expenditures

by

all

payers

on

specific

categories

of

care:

inpa-tient,

outpatient,

prescription

drugs,

and

other

(which

includes

dental,

vision,

home

health

care

services,

and

medical

equipment

but

excludes

spending

on

over-the-counter

medications).

Medi-cal

expenditures

by

source

of

payment

are

collected

directly

from

households

as

well

as

from

the

household’s

medical

care

providers

for

every

medical

event.

In

addition,

MEPS

respondents

are

asked

whether

their

medical

visits

or

other

events

are

related

to

any

specific

medical

conditions.

These

responses

are

then

profession-ally

coded

using

the

International

Classification

of

Diseases

,

Ninth

Revision

(ICD-9),

and

subsequently

collapsed

to

into

259

clinically

relevant

medical

conditions

using

the

Clinical

Classification

System

(CCS)

developed

by

the

Agency

for

Healthcare

Research

and

Quality

(AHRQ,

2007).

MEPS

data

are

collected

through

a

stratified

multi-stage

prob-ability

design,

which

we

account

for

in

the

calculation

of

the

standard

errors

for

our

marginal

effects.

In

particular,

we

use

the

method

of

balanced

repeated

replications

to

estimate

standard

errors

in

our

2PM

and

the

method

of

bootstrapping

with

500

9

_As

_a

_robustness

_check,

_we

_also

_estimated

_models

_that

_include

_controls

_for

_health

insurance

status

(private

or

public

insurance),

employer

size

(indicators

of

whether

the

firm

contained

less

than

25,

25–100,

101–500,

or

over

500

employees),

whether

the

individual

belonged

to

a

union,

whether

the

individual

was

married,

and

net

income

per

adult

equivalent

(total

household

income

minus

health

insurance

pre-miums

divided

by

the

square

root

of

household

size).

Including

these

additional

regressors

has

little

effect

on

the

estimated

impact

of

BMI

and

obesity

on

medical

expenditures.

10

_The

_exception

_to

_this

_is

_when

_all

_adult

_members

_of

_the

_household

_are

_present

during

the

interview,

in

which

case

each

adult

self-reports

their

height

and

weight.

Table

1

Descriptive

statistics

for

sample

of

men

with

biological

children

(

N

=

9852).

Mean

S.D.

Min

Max

Medical

expenditures

>

0

0.79

0.41

0

1

Total

medical

expenditures

1999

5406

0

212,681

Third

party

med.

expenditures

>

0

0.72

0.45

0

1

Third

party

medical

expenditures

1577

4970

0

197,501

Body

mass

index

28.17

4.88

14.98

59.30

Body

mass

index

squared

817

308

224

3516

Obesity

(BMI

≥

30)

0.28

0.45

0

1

White

0.72

0.45

0

1

Hispanic

0.14

0.35

0

1

Black

0.09

0.28

0

1

Other

race

0.05

0.23

0

1

Age

is

20–34

0.06

0.24

0

1

Age

is

35–44

0.43

0.49

0

1

Age

is

45–54

0.43

0.49

0

1

Age

is

55–64

0.09

0.28

0

1

No.

high

school

diploma

0.16

0.37

0

1

High

school

graduate

0.34

0.47

0

1

Some

college

0.21

0.41

0

1

Bachelor’s

degree

or

higher

0.29

0.45

0

1

Northeastern

census

region

0.19

0.39

0

1

Midwestern

census

region

0.24

0.43

0

1

Southern

census

region

0.34

0.47

0

1

Western

census

region

0.23

0.42

0

1

Residence

in

MSA

0.81

0.40

0

1

No.

HH

members

0–5

0.18

0.48

0

5

No.

HH

members

6–17

1.74

1.02

0

9

No.

HH

members

18–64

2.30

0.73

1

9

No.

HH

members

65+

0.02

0.17

0

3

Survey

info.

is

self-reported

0.25

0.43

0

1

Not

employed

0.09

0.28

0

1

Oldest

child

is

female

0.47

0.50

0

1

Oldest

child

age

in

months

191.44

31.20

132.00

240.00

Oldest

child

BMI

22.49

5.10

7.60

109.19

Oldest

child

BMI

squared

532

330

58

11,923

Oldest

child

BMI

cubed

13,504

26,148

439

1,301,949

Oldest

child

is

obese

0.12

0.33

0

1

Year

is

2000

0.16

0.36

0

1

Year

is

2001

0.17

0.37

0

1

Year

is

2002

0.17

0.38

0

1

Year

is

2003

0.17

0.37

0

1

Year

is

2004

0.16

0.37

0

1

Year

is

2005

0.17

0.37

0

1

Notes:

Data:

MEPS

2000–2005.

All

means

are

weighted

and

monetary

values

expressed

in

2005

USD.

replications

in

the

IV

QTE

models.

Both

methods

account

for

clustering

at

the

PSU-level,

stratification,

and

weighting.

4.

Results

4.1.

Summary

statistics

Descriptive

statistics

for

the

main

set

of

variables

used

in

our

empirical

analysis

are

contained

in

Table

1

for

men

and

Table

2

for

women.

(The

samples

are

limited

to

adults

with

biological

children,

as

they

are

the

only

MEPS

respondents

for

whom

we

can

estimate

the

IV

model.)

Among

men,

79%

incur

some

medical

expenditures

in

the

survey

year,

and

the

unconditional

average

medical

expen-ditures

in

that

year

was

$1999

(which

includes

zeros

for

those

with

no

expenditures)

in

2005

dollars.

Among

women,

88%

incurred

some

medical

expenditures,

and

the

unconditional

average

medical

expenditures

in

that

year

was

$2617.

These

expenditures

are

lower

than

those

for

the

comparable

population

of

all

men

and

women

(i.e.

those

both

with

and

without

biological

children)

(AHRQ,

2010).

For

men

(women),

approximately

72%

(80%)

of

expenditures

are

covered

by

third

party

payers.